blob: a3fe6faba3d1373f7ab94d991b4b59f3fadd68d6 [file] [log] [blame]
Evan Chengb1290a62008-10-02 18:29:27 +00001//===------ RegAllocPBQP.cpp ---- PBQP Register Allocator -------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
Misha Brukman2a835f92009-01-08 15:50:22 +00009//
Evan Chengb1290a62008-10-02 18:29:27 +000010// This file contains a Partitioned Boolean Quadratic Programming (PBQP) based
11// register allocator for LLVM. This allocator works by constructing a PBQP
12// problem representing the register allocation problem under consideration,
13// solving this using a PBQP solver, and mapping the solution back to a
14// register assignment. If any variables are selected for spilling then spill
Misha Brukman2a835f92009-01-08 15:50:22 +000015// code is inserted and the process repeated.
Evan Chengb1290a62008-10-02 18:29:27 +000016//
17// The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
18// for register allocation. For more information on PBQP for register
Misha Brukmance07e992009-01-08 16:40:25 +000019// allocation, see the following papers:
Evan Chengb1290a62008-10-02 18:29:27 +000020//
21// (1) Hames, L. and Scholz, B. 2006. Nearly optimal register allocation with
22// PBQP. In Proceedings of the 7th Joint Modular Languages Conference
23// (JMLC'06). LNCS, vol. 4228. Springer, New York, NY, USA. 346-361.
24//
25// (2) Scholz, B., Eckstein, E. 2002. Register allocation for irregular
26// architectures. In Proceedings of the Joint Conference on Languages,
27// Compilers and Tools for Embedded Systems (LCTES'02), ACM Press, New York,
28// NY, USA, 139-148.
Misha Brukman2a835f92009-01-08 15:50:22 +000029//
Evan Chengb1290a62008-10-02 18:29:27 +000030//===----------------------------------------------------------------------===//
31
Evan Chengb1290a62008-10-02 18:29:27 +000032#define DEBUG_TYPE "regalloc"
33
Lang Hames54cc2ef2010-07-19 15:22:28 +000034#include "RenderMachineFunction.h"
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +000035#include "Spiller.h"
Evan Chengb1290a62008-10-02 18:29:27 +000036#include "VirtRegMap.h"
Rafael Espindolafdf16ca2011-06-26 21:41:06 +000037#include "RegisterCoalescer.h"
Lang Hames20df03c2012-03-26 23:07:23 +000038#include "llvm/Module.h"
Lang Hames9ad7e072011-12-06 01:45:57 +000039#include "llvm/Analysis/AliasAnalysis.h"
Lang Hamesa937f222009-12-14 06:49:42 +000040#include "llvm/CodeGen/CalcSpillWeights.h"
Evan Chengb1290a62008-10-02 18:29:27 +000041#include "llvm/CodeGen/LiveIntervalAnalysis.h"
Pete Cooper789d5d82012-04-02 22:44:18 +000042#include "llvm/CodeGen/LiveRangeEdit.h"
Lang Hames27601ef2008-11-16 12:12:54 +000043#include "llvm/CodeGen/LiveStackAnalysis.h"
Lang Hameseb6c8f52010-09-18 09:07:10 +000044#include "llvm/CodeGen/RegAllocPBQP.h"
Lang Hames9ad7e072011-12-06 01:45:57 +000045#include "llvm/CodeGen/MachineDominators.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000046#include "llvm/CodeGen/MachineFunctionPass.h"
Evan Chengb1290a62008-10-02 18:29:27 +000047#include "llvm/CodeGen/MachineLoopInfo.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000048#include "llvm/CodeGen/MachineRegisterInfo.h"
Lang Hameseb6c8f52010-09-18 09:07:10 +000049#include "llvm/CodeGen/PBQP/HeuristicSolver.h"
50#include "llvm/CodeGen/PBQP/Graph.h"
51#include "llvm/CodeGen/PBQP/Heuristics/Briggs.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000052#include "llvm/CodeGen/RegAllocRegistry.h"
Evan Chengb1290a62008-10-02 18:29:27 +000053#include "llvm/Support/Debug.h"
Daniel Dunbarce63ffb2009-07-25 00:23:56 +000054#include "llvm/Support/raw_ostream.h"
Misha Brukman2a835f92009-01-08 15:50:22 +000055#include "llvm/Target/TargetInstrInfo.h"
56#include "llvm/Target/TargetMachine.h"
57#include <limits>
Misha Brukman2a835f92009-01-08 15:50:22 +000058#include <memory>
Evan Chengb1290a62008-10-02 18:29:27 +000059#include <set>
Lang Hames20df03c2012-03-26 23:07:23 +000060#include <sstream>
Evan Chengb1290a62008-10-02 18:29:27 +000061#include <vector>
Evan Chengb1290a62008-10-02 18:29:27 +000062
Lang Hamesf70e7cc2010-09-23 04:28:54 +000063using namespace llvm;
Lang Hameseb6c8f52010-09-18 09:07:10 +000064
Evan Chengb1290a62008-10-02 18:29:27 +000065static RegisterRegAlloc
Duncan Sands1aecd152010-02-18 14:10:41 +000066registerPBQPRepAlloc("pbqp", "PBQP register allocator",
Lang Hamesf70e7cc2010-09-23 04:28:54 +000067 createDefaultPBQPRegisterAllocator);
Evan Chengb1290a62008-10-02 18:29:27 +000068
Lang Hames8481e3b2009-08-19 01:36:14 +000069static cl::opt<bool>
70pbqpCoalescing("pbqp-coalescing",
Lang Hames030c4bf2010-01-26 04:49:58 +000071 cl::desc("Attempt coalescing during PBQP register allocation."),
72 cl::init(false), cl::Hidden);
Lang Hames8481e3b2009-08-19 01:36:14 +000073
Lang Hames20df03c2012-03-26 23:07:23 +000074#ifndef NDEBUG
75static cl::opt<bool>
76pbqpDumpGraphs("pbqp-dump-graphs",
77 cl::desc("Dump graphs for each function/round in the compilation unit."),
78 cl::init(false), cl::Hidden);
79#endif
80
Lang Hamesf70e7cc2010-09-23 04:28:54 +000081namespace {
82
83///
84/// PBQP based allocators solve the register allocation problem by mapping
85/// register allocation problems to Partitioned Boolean Quadratic
86/// Programming problems.
87class RegAllocPBQP : public MachineFunctionPass {
88public:
89
90 static char ID;
91
92 /// Construct a PBQP register allocator.
Lang Hames8d857662011-06-17 07:09:01 +000093 RegAllocPBQP(std::auto_ptr<PBQPBuilder> b, char *cPassID=0)
94 : MachineFunctionPass(ID), builder(b), customPassID(cPassID) {
Owen Anderson081c34b2010-10-19 17:21:58 +000095 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
96 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
Owen Anderson081c34b2010-10-19 17:21:58 +000097 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
98 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
99 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
Owen Anderson081c34b2010-10-19 17:21:58 +0000100 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
101 initializeRenderMachineFunctionPass(*PassRegistry::getPassRegistry());
102 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000103
104 /// Return the pass name.
105 virtual const char* getPassName() const {
106 return "PBQP Register Allocator";
107 }
108
109 /// PBQP analysis usage.
110 virtual void getAnalysisUsage(AnalysisUsage &au) const;
111
112 /// Perform register allocation
113 virtual bool runOnMachineFunction(MachineFunction &MF);
114
115private:
116
117 typedef std::map<const LiveInterval*, unsigned> LI2NodeMap;
118 typedef std::vector<const LiveInterval*> Node2LIMap;
119 typedef std::vector<unsigned> AllowedSet;
120 typedef std::vector<AllowedSet> AllowedSetMap;
121 typedef std::pair<unsigned, unsigned> RegPair;
122 typedef std::map<RegPair, PBQP::PBQPNum> CoalesceMap;
123 typedef std::vector<PBQP::Graph::NodeItr> NodeVector;
124 typedef std::set<unsigned> RegSet;
125
126
127 std::auto_ptr<PBQPBuilder> builder;
128
Lang Hames8d857662011-06-17 07:09:01 +0000129 char *customPassID;
130
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000131 MachineFunction *mf;
132 const TargetMachine *tm;
133 const TargetRegisterInfo *tri;
134 const TargetInstrInfo *tii;
135 const MachineLoopInfo *loopInfo;
136 MachineRegisterInfo *mri;
137 RenderMachineFunction *rmf;
138
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000139 std::auto_ptr<Spiller> spiller;
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000140 LiveIntervals *lis;
141 LiveStacks *lss;
142 VirtRegMap *vrm;
143
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000144 RegSet vregsToAlloc, emptyIntervalVRegs;
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000145
146 /// \brief Finds the initial set of vreg intervals to allocate.
147 void findVRegIntervalsToAlloc();
148
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000149 /// \brief Given a solved PBQP problem maps this solution back to a register
150 /// assignment.
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000151 bool mapPBQPToRegAlloc(const PBQPRAProblem &problem,
152 const PBQP::Solution &solution);
153
154 /// \brief Postprocessing before final spilling. Sets basic block "live in"
155 /// variables.
156 void finalizeAlloc() const;
157
158};
159
Lang Hameseb6c8f52010-09-18 09:07:10 +0000160char RegAllocPBQP::ID = 0;
Evan Chengb1290a62008-10-02 18:29:27 +0000161
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000162} // End anonymous namespace.
163
Lang Hameseb6c8f52010-09-18 09:07:10 +0000164unsigned PBQPRAProblem::getVRegForNode(PBQP::Graph::ConstNodeItr node) const {
165 Node2VReg::const_iterator vregItr = node2VReg.find(node);
166 assert(vregItr != node2VReg.end() && "No vreg for node.");
167 return vregItr->second;
168}
Evan Chengb1290a62008-10-02 18:29:27 +0000169
Lang Hameseb6c8f52010-09-18 09:07:10 +0000170PBQP::Graph::NodeItr PBQPRAProblem::getNodeForVReg(unsigned vreg) const {
171 VReg2Node::const_iterator nodeItr = vreg2Node.find(vreg);
172 assert(nodeItr != vreg2Node.end() && "No node for vreg.");
173 return nodeItr->second;
Andrew Trick16f72dd2012-02-10 04:10:26 +0000174
Lang Hameseb6c8f52010-09-18 09:07:10 +0000175}
Daniel Dunbara279bc32009-09-20 02:20:51 +0000176
Lang Hameseb6c8f52010-09-18 09:07:10 +0000177const PBQPRAProblem::AllowedSet&
178 PBQPRAProblem::getAllowedSet(unsigned vreg) const {
179 AllowedSetMap::const_iterator allowedSetItr = allowedSets.find(vreg);
180 assert(allowedSetItr != allowedSets.end() && "No pregs for vreg.");
181 const AllowedSet &allowedSet = allowedSetItr->second;
182 return allowedSet;
183}
Evan Chengb1290a62008-10-02 18:29:27 +0000184
Lang Hameseb6c8f52010-09-18 09:07:10 +0000185unsigned PBQPRAProblem::getPRegForOption(unsigned vreg, unsigned option) const {
186 assert(isPRegOption(vreg, option) && "Not a preg option.");
187
188 const AllowedSet& allowedSet = getAllowedSet(vreg);
189 assert(option <= allowedSet.size() && "Option outside allowed set.");
190 return allowedSet[option - 1];
191}
192
Lang Hamese9c93562010-09-21 13:19:36 +0000193std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(MachineFunction *mf,
194 const LiveIntervals *lis,
195 const MachineLoopInfo *loopInfo,
196 const RegSet &vregs) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000197
198 typedef std::vector<const LiveInterval*> LIVector;
Jakob Stoklund Olesen3b30bca2012-06-20 22:32:05 +0000199 LiveIntervals *LIS = const_cast<LiveIntervals*>(lis);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000200 MachineRegisterInfo *mri = &mf->getRegInfo();
Andrew Trick16f72dd2012-02-10 04:10:26 +0000201 const TargetRegisterInfo *tri = mf->getTarget().getRegisterInfo();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000202
203 std::auto_ptr<PBQPRAProblem> p(new PBQPRAProblem());
204 PBQP::Graph &g = p->getGraph();
205 RegSet pregs;
206
207 // Collect the set of preg intervals, record that they're used in the MF.
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000208 for (unsigned Reg = 1, e = tri->getNumRegs(); Reg != e; ++Reg) {
Jakob Stoklund Olesen3b30bca2012-06-20 22:32:05 +0000209 if (mri->def_empty(Reg))
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000210 continue;
211 pregs.insert(Reg);
212 mri->setPhysRegUsed(Reg);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000213 }
Evan Chengb1290a62008-10-02 18:29:27 +0000214
Lang Hameseb6c8f52010-09-18 09:07:10 +0000215 BitVector reservedRegs = tri->getReservedRegs(*mf);
Evan Chengb1290a62008-10-02 18:29:27 +0000216
Andrew Trick16f72dd2012-02-10 04:10:26 +0000217 // Iterate over vregs.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000218 for (RegSet::const_iterator vregItr = vregs.begin(), vregEnd = vregs.end();
219 vregItr != vregEnd; ++vregItr) {
220 unsigned vreg = *vregItr;
221 const TargetRegisterClass *trc = mri->getRegClass(vreg);
Jakob Stoklund Olesen3b30bca2012-06-20 22:32:05 +0000222 LiveInterval *vregLI = &LIS->getInterval(vreg);
223
224 // Record any overlaps with regmask operands.
225 BitVector regMaskOverlaps(tri->getNumRegs());
226 LIS->checkRegMaskInterference(*vregLI, regMaskOverlaps);
Evan Chengb1290a62008-10-02 18:29:27 +0000227
Lang Hameseb6c8f52010-09-18 09:07:10 +0000228 // Compute an initial allowed set for the current vreg.
229 typedef std::vector<unsigned> VRAllowed;
230 VRAllowed vrAllowed;
Craig Topperb6632ba2012-03-04 10:16:38 +0000231 ArrayRef<uint16_t> rawOrder = trc->getRawAllocationOrder(*mf);
Jakob Stoklund Olesen714c0eb2011-06-16 20:37:45 +0000232 for (unsigned i = 0; i != rawOrder.size(); ++i) {
233 unsigned preg = rawOrder[i];
Jakob Stoklund Olesen3b30bca2012-06-20 22:32:05 +0000234 if (reservedRegs.test(preg))
235 continue;
236
237 // vregLI crosses a regmask operand that clobbers preg.
238 if (!regMaskOverlaps.empty() && !regMaskOverlaps.test(preg))
239 continue;
240
241 // vregLI overlaps fixed regunit interference.
242 if (LIS->trackingRegUnits()) {
243 bool Interference = false;
244 for (MCRegUnitIterator Units(preg, tri); Units.isValid(); ++Units) {
245 if (vregLI->overlaps(LIS->getRegUnit(*Units))) {
246 Interference = true;
247 break;
248 }
249 }
250 if (Interference)
251 continue;
Lang Hamesd0f6f012010-07-17 06:31:41 +0000252 }
Jakob Stoklund Olesen3b30bca2012-06-20 22:32:05 +0000253
254 // preg is usable for this virtual register.
255 vrAllowed.push_back(preg);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000256 }
Lang Hamesd0f6f012010-07-17 06:31:41 +0000257
Lang Hamesf1113ef2012-03-23 17:33:42 +0000258 RegSet overlappingPRegs;
259
260 // Record physical registers whose ranges overlap.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000261 for (RegSet::const_iterator pregItr = pregs.begin(),
262 pregEnd = pregs.end();
263 pregItr != pregEnd; ++pregItr) {
264 unsigned preg = *pregItr;
Jakob Stoklund Olesen3b30bca2012-06-20 22:32:05 +0000265 if (!LIS->hasInterval(preg))
266 continue;
267 const LiveInterval *pregLI = &LIS->getInterval(preg);
Lang Hames27601ef2008-11-16 12:12:54 +0000268
Lang Hames5e77f4b2010-11-12 05:47:21 +0000269 if (pregLI->empty()) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000270 continue;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000271 }
Evan Chengb1290a62008-10-02 18:29:27 +0000272
Lang Hamesf1113ef2012-03-23 17:33:42 +0000273 if (vregLI->overlaps(*pregLI))
274 overlappingPRegs.insert(preg);
275 }
276
Lang Hamesf1113ef2012-03-23 17:33:42 +0000277 for (RegSet::const_iterator pregItr = overlappingPRegs.begin(),
278 pregEnd = overlappingPRegs.end();
279 pregItr != pregEnd; ++pregItr) {
280 unsigned preg = *pregItr;
Lang Hames030c4bf2010-01-26 04:49:58 +0000281
Lang Hameseb6c8f52010-09-18 09:07:10 +0000282 // Remove the register from the allowed set.
283 VRAllowed::iterator eraseItr =
284 std::find(vrAllowed.begin(), vrAllowed.end(), preg);
Evan Chengb1290a62008-10-02 18:29:27 +0000285
Lang Hameseb6c8f52010-09-18 09:07:10 +0000286 if (eraseItr != vrAllowed.end()) {
287 vrAllowed.erase(eraseItr);
288 }
Evan Chengb1290a62008-10-02 18:29:27 +0000289
Lang Hameseb6c8f52010-09-18 09:07:10 +0000290 // Also remove any aliases.
Jakob Stoklund Olesen396618b2012-06-01 23:28:30 +0000291 for (MCRegAliasIterator AI(preg, tri, false); AI.isValid(); ++AI) {
292 VRAllowed::iterator eraseItr =
293 std::find(vrAllowed.begin(), vrAllowed.end(), *AI);
294 if (eraseItr != vrAllowed.end()) {
295 vrAllowed.erase(eraseItr);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000296 }
297 }
298 }
Misha Brukman2a835f92009-01-08 15:50:22 +0000299
Lang Hameseb6c8f52010-09-18 09:07:10 +0000300 // Construct the node.
Andrew Trick16f72dd2012-02-10 04:10:26 +0000301 PBQP::Graph::NodeItr node =
Lang Hameseb6c8f52010-09-18 09:07:10 +0000302 g.addNode(PBQP::Vector(vrAllowed.size() + 1, 0));
Evan Chengb1290a62008-10-02 18:29:27 +0000303
Lang Hameseb6c8f52010-09-18 09:07:10 +0000304 // Record the mapping and allowed set in the problem.
305 p->recordVReg(vreg, node, vrAllowed.begin(), vrAllowed.end());
Evan Chengb1290a62008-10-02 18:29:27 +0000306
Lang Hameseb6c8f52010-09-18 09:07:10 +0000307 PBQP::PBQPNum spillCost = (vregLI->weight != 0.0) ?
308 vregLI->weight : std::numeric_limits<PBQP::PBQPNum>::min();
Evan Chengb1290a62008-10-02 18:29:27 +0000309
Lang Hameseb6c8f52010-09-18 09:07:10 +0000310 addSpillCosts(g.getNodeCosts(node), spillCost);
311 }
Evan Chengb1290a62008-10-02 18:29:27 +0000312
Lang Hames481630d2010-09-18 09:49:08 +0000313 for (RegSet::const_iterator vr1Itr = vregs.begin(), vrEnd = vregs.end();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000314 vr1Itr != vrEnd; ++vr1Itr) {
315 unsigned vr1 = *vr1Itr;
316 const LiveInterval &l1 = lis->getInterval(vr1);
317 const PBQPRAProblem::AllowedSet &vr1Allowed = p->getAllowedSet(vr1);
Evan Chengb1290a62008-10-02 18:29:27 +0000318
Benjamin Kramer9e8d1f92010-09-18 14:41:26 +0000319 for (RegSet::const_iterator vr2Itr = llvm::next(vr1Itr);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000320 vr2Itr != vrEnd; ++vr2Itr) {
321 unsigned vr2 = *vr2Itr;
322 const LiveInterval &l2 = lis->getInterval(vr2);
323 const PBQPRAProblem::AllowedSet &vr2Allowed = p->getAllowedSet(vr2);
Evan Chengb1290a62008-10-02 18:29:27 +0000324
Lang Hameseb6c8f52010-09-18 09:07:10 +0000325 assert(!l2.empty() && "Empty interval in vreg set?");
326 if (l1.overlaps(l2)) {
327 PBQP::Graph::EdgeItr edge =
328 g.addEdge(p->getNodeForVReg(vr1), p->getNodeForVReg(vr2),
329 PBQP::Matrix(vr1Allowed.size()+1, vr2Allowed.size()+1, 0));
Lang Hames27601ef2008-11-16 12:12:54 +0000330
Lang Hameseb6c8f52010-09-18 09:07:10 +0000331 addInterferenceCosts(g.getEdgeCosts(edge), vr1Allowed, vr2Allowed, tri);
332 }
333 }
334 }
Evan Chengb1290a62008-10-02 18:29:27 +0000335
Lang Hameseb6c8f52010-09-18 09:07:10 +0000336 return p;
337}
Lang Hames27601ef2008-11-16 12:12:54 +0000338
Lang Hameseb6c8f52010-09-18 09:07:10 +0000339void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
340 PBQP::PBQPNum spillCost) {
341 costVec[0] = spillCost;
342}
Evan Chengb1290a62008-10-02 18:29:27 +0000343
Lang Hamese9c93562010-09-21 13:19:36 +0000344void PBQPBuilder::addInterferenceCosts(
345 PBQP::Matrix &costMat,
346 const PBQPRAProblem::AllowedSet &vr1Allowed,
347 const PBQPRAProblem::AllowedSet &vr2Allowed,
348 const TargetRegisterInfo *tri) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000349 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Matrix height mismatch.");
350 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Matrix width mismatch.");
351
Lang Hames5e77f4b2010-11-12 05:47:21 +0000352 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000353 unsigned preg1 = vr1Allowed[i];
354
Lang Hames5e77f4b2010-11-12 05:47:21 +0000355 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000356 unsigned preg2 = vr2Allowed[j];
357
358 if (tri->regsOverlap(preg1, preg2)) {
359 costMat[i + 1][j + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
360 }
361 }
362 }
Evan Chengb1290a62008-10-02 18:29:27 +0000363}
364
Lang Hamese9c93562010-09-21 13:19:36 +0000365std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
366 MachineFunction *mf,
367 const LiveIntervals *lis,
368 const MachineLoopInfo *loopInfo,
369 const RegSet &vregs) {
370
371 std::auto_ptr<PBQPRAProblem> p = PBQPBuilder::build(mf, lis, loopInfo, vregs);
372 PBQP::Graph &g = p->getGraph();
373
374 const TargetMachine &tm = mf->getTarget();
Benjamin Kramera7542d52012-06-06 18:25:08 +0000375 CoalescerPair cp(*tm.getRegisterInfo());
Lang Hamese9c93562010-09-21 13:19:36 +0000376
377 // Scan the machine function and add a coalescing cost whenever CoalescerPair
378 // gives the Ok.
379 for (MachineFunction::const_iterator mbbItr = mf->begin(),
380 mbbEnd = mf->end();
381 mbbItr != mbbEnd; ++mbbItr) {
382 const MachineBasicBlock *mbb = &*mbbItr;
383
384 for (MachineBasicBlock::const_iterator miItr = mbb->begin(),
385 miEnd = mbb->end();
386 miItr != miEnd; ++miItr) {
387 const MachineInstr *mi = &*miItr;
388
Lang Hames5e77f4b2010-11-12 05:47:21 +0000389 if (!cp.setRegisters(mi)) {
Lang Hamese9c93562010-09-21 13:19:36 +0000390 continue; // Not coalescable.
Lang Hames5e77f4b2010-11-12 05:47:21 +0000391 }
Lang Hamese9c93562010-09-21 13:19:36 +0000392
Lang Hames5e77f4b2010-11-12 05:47:21 +0000393 if (cp.getSrcReg() == cp.getDstReg()) {
Lang Hamese9c93562010-09-21 13:19:36 +0000394 continue; // Already coalesced.
Lang Hames5e77f4b2010-11-12 05:47:21 +0000395 }
Lang Hamese9c93562010-09-21 13:19:36 +0000396
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000397 unsigned dst = cp.getDstReg(),
398 src = cp.getSrcReg();
Lang Hamese9c93562010-09-21 13:19:36 +0000399
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000400 const float copyFactor = 0.5; // Cost of copy relative to load. Current
401 // value plucked randomly out of the air.
Andrew Trick16f72dd2012-02-10 04:10:26 +0000402
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000403 PBQP::PBQPNum cBenefit =
404 copyFactor * LiveIntervals::getSpillWeight(false, true,
405 loopInfo->getLoopDepth(mbb));
Lang Hamese9c93562010-09-21 13:19:36 +0000406
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000407 if (cp.isPhys()) {
Lang Hames5e77f4b2010-11-12 05:47:21 +0000408 if (!lis->isAllocatable(dst)) {
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000409 continue;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000410 }
Lang Hamese9c93562010-09-21 13:19:36 +0000411
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000412 const PBQPRAProblem::AllowedSet &allowed = p->getAllowedSet(src);
Andrew Trick16f72dd2012-02-10 04:10:26 +0000413 unsigned pregOpt = 0;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000414 while (pregOpt < allowed.size() && allowed[pregOpt] != dst) {
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000415 ++pregOpt;
Lang Hames5e77f4b2010-11-12 05:47:21 +0000416 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000417 if (pregOpt < allowed.size()) {
418 ++pregOpt; // +1 to account for spill option.
419 PBQP::Graph::NodeItr node = p->getNodeForVReg(src);
420 addPhysRegCoalesce(g.getNodeCosts(node), pregOpt, cBenefit);
Lang Hamese9c93562010-09-21 13:19:36 +0000421 }
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000422 } else {
423 const PBQPRAProblem::AllowedSet *allowed1 = &p->getAllowedSet(dst);
424 const PBQPRAProblem::AllowedSet *allowed2 = &p->getAllowedSet(src);
425 PBQP::Graph::NodeItr node1 = p->getNodeForVReg(dst);
426 PBQP::Graph::NodeItr node2 = p->getNodeForVReg(src);
427 PBQP::Graph::EdgeItr edge = g.findEdge(node1, node2);
428 if (edge == g.edgesEnd()) {
429 edge = g.addEdge(node1, node2, PBQP::Matrix(allowed1->size() + 1,
430 allowed2->size() + 1,
431 0));
432 } else {
433 if (g.getEdgeNode1(edge) == node2) {
434 std::swap(node1, node2);
435 std::swap(allowed1, allowed2);
436 }
437 }
Andrew Trick16f72dd2012-02-10 04:10:26 +0000438
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000439 addVirtRegCoalesce(g.getEdgeCosts(edge), *allowed1, *allowed2,
440 cBenefit);
Lang Hamese9c93562010-09-21 13:19:36 +0000441 }
442 }
443 }
444
445 return p;
446}
447
Lang Hamese9c93562010-09-21 13:19:36 +0000448void PBQPBuilderWithCoalescing::addPhysRegCoalesce(PBQP::Vector &costVec,
449 unsigned pregOption,
450 PBQP::PBQPNum benefit) {
451 costVec[pregOption] += -benefit;
452}
453
454void PBQPBuilderWithCoalescing::addVirtRegCoalesce(
455 PBQP::Matrix &costMat,
456 const PBQPRAProblem::AllowedSet &vr1Allowed,
457 const PBQPRAProblem::AllowedSet &vr2Allowed,
458 PBQP::PBQPNum benefit) {
459
460 assert(costMat.getRows() == vr1Allowed.size() + 1 && "Size mismatch.");
461 assert(costMat.getCols() == vr2Allowed.size() + 1 && "Size mismatch.");
462
Lang Hames5e77f4b2010-11-12 05:47:21 +0000463 for (unsigned i = 0; i != vr1Allowed.size(); ++i) {
Lang Hamese9c93562010-09-21 13:19:36 +0000464 unsigned preg1 = vr1Allowed[i];
Lang Hames5e77f4b2010-11-12 05:47:21 +0000465 for (unsigned j = 0; j != vr2Allowed.size(); ++j) {
Lang Hamese9c93562010-09-21 13:19:36 +0000466 unsigned preg2 = vr2Allowed[j];
467
468 if (preg1 == preg2) {
469 costMat[i + 1][j + 1] += -benefit;
Andrew Trick16f72dd2012-02-10 04:10:26 +0000470 }
Lang Hamese9c93562010-09-21 13:19:36 +0000471 }
472 }
473}
Evan Chengb1290a62008-10-02 18:29:27 +0000474
Lang Hameseb6c8f52010-09-18 09:07:10 +0000475
476void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
Lang Hames9ad7e072011-12-06 01:45:57 +0000477 au.setPreservesCFG();
478 au.addRequired<AliasAnalysis>();
479 au.addPreserved<AliasAnalysis>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000480 au.addRequired<SlotIndexes>();
481 au.addPreserved<SlotIndexes>();
482 au.addRequired<LiveIntervals>();
483 //au.addRequiredID(SplitCriticalEdgesID);
Lang Hames8d857662011-06-17 07:09:01 +0000484 if (customPassID)
485 au.addRequiredID(*customPassID);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000486 au.addRequired<CalculateSpillWeights>();
487 au.addRequired<LiveStacks>();
488 au.addPreserved<LiveStacks>();
Lang Hames9ad7e072011-12-06 01:45:57 +0000489 au.addRequired<MachineDominatorTree>();
490 au.addPreserved<MachineDominatorTree>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000491 au.addRequired<MachineLoopInfo>();
492 au.addPreserved<MachineLoopInfo>();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000493 au.addRequired<VirtRegMap>();
494 au.addRequired<RenderMachineFunction>();
495 MachineFunctionPass::getAnalysisUsage(au);
496}
497
Lang Hameseb6c8f52010-09-18 09:07:10 +0000498void RegAllocPBQP::findVRegIntervalsToAlloc() {
Lang Hames27601ef2008-11-16 12:12:54 +0000499
500 // Iterate over all live ranges.
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000501 for (unsigned i = 0, e = mri->getNumVirtRegs(); i != e; ++i) {
502 unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
503 if (mri->reg_nodbg_empty(Reg))
Lang Hames27601ef2008-11-16 12:12:54 +0000504 continue;
Jakob Stoklund Olesend67582e2012-06-20 21:25:05 +0000505 LiveInterval *li = &lis->getInterval(Reg);
Lang Hames27601ef2008-11-16 12:12:54 +0000506
507 // If this live interval is non-empty we will use pbqp to allocate it.
508 // Empty intervals we allocate in a simple post-processing stage in
509 // finalizeAlloc.
510 if (!li->empty()) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000511 vregsToAlloc.insert(li->reg);
Lang Hames5e77f4b2010-11-12 05:47:21 +0000512 } else {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000513 emptyIntervalVRegs.insert(li->reg);
Lang Hames27601ef2008-11-16 12:12:54 +0000514 }
515 }
Evan Chengb1290a62008-10-02 18:29:27 +0000516}
517
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000518bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAProblem &problem,
519 const PBQP::Solution &solution) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000520 // Set to true if we have any spills
521 bool anotherRoundNeeded = false;
522
523 // Clear the existing allocation.
524 vrm->clearAllVirt();
525
526 const PBQP::Graph &g = problem.getGraph();
527 // Iterate over the nodes mapping the PBQP solution to a register
528 // assignment.
529 for (PBQP::Graph::ConstNodeItr node = g.nodesBegin(),
530 nodeEnd = g.nodesEnd();
531 node != nodeEnd; ++node) {
532 unsigned vreg = problem.getVRegForNode(node);
533 unsigned alloc = solution.getSelection(node);
534
535 if (problem.isPRegOption(vreg, alloc)) {
Andrew Trick16f72dd2012-02-10 04:10:26 +0000536 unsigned preg = problem.getPRegForOption(vreg, alloc);
Patrik Hägglundd7693872012-05-23 12:12:58 +0000537 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> "
538 << tri->getName(preg) << "\n");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000539 assert(preg != 0 && "Invalid preg selected.");
Andrew Trick16f72dd2012-02-10 04:10:26 +0000540 vrm->assignVirt2Phys(vreg, preg);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000541 } else if (problem.isSpillOption(vreg, alloc)) {
542 vregsToAlloc.erase(vreg);
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000543 SmallVector<LiveInterval*, 8> newSpills;
Jakob Stoklund Olesen20942dc2012-05-19 05:25:46 +0000544 LiveRangeEdit LRE(&lis->getInterval(vreg), newSpills, *mf, *lis, vrm);
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000545 spiller->spill(LRE);
Lang Hameseb6c8f52010-09-18 09:07:10 +0000546
Patrik Hägglundd7693872012-05-23 12:12:58 +0000547 DEBUG(dbgs() << "VREG " << PrintReg(vreg, tri) << " -> SPILLED (Cost: "
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000548 << LRE.getParent().weight << ", New vregs: ");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000549
550 // Copy any newly inserted live intervals into the list of regs to
551 // allocate.
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000552 for (LiveRangeEdit::iterator itr = LRE.begin(), end = LRE.end();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000553 itr != end; ++itr) {
554 assert(!(*itr)->empty() && "Empty spill range.");
Patrik Hägglundd7693872012-05-23 12:12:58 +0000555 DEBUG(dbgs() << PrintReg((*itr)->reg, tri) << " ");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000556 vregsToAlloc.insert((*itr)->reg);
557 }
558
559 DEBUG(dbgs() << ")\n");
560
561 // We need another round if spill intervals were added.
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000562 anotherRoundNeeded |= !LRE.empty();
Lang Hameseb6c8f52010-09-18 09:07:10 +0000563 } else {
Craig Topper5e25ee82012-02-05 08:31:47 +0000564 llvm_unreachable("Unknown allocation option.");
Lang Hameseb6c8f52010-09-18 09:07:10 +0000565 }
566 }
567
568 return !anotherRoundNeeded;
569}
570
571
572void RegAllocPBQP::finalizeAlloc() const {
Lang Hames27601ef2008-11-16 12:12:54 +0000573 typedef LiveIntervals::iterator LIIterator;
574 typedef LiveInterval::Ranges::const_iterator LRIterator;
575
576 // First allocate registers for the empty intervals.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000577 for (RegSet::const_iterator
578 itr = emptyIntervalVRegs.begin(), end = emptyIntervalVRegs.end();
Lang Hames27601ef2008-11-16 12:12:54 +0000579 itr != end; ++itr) {
Lang Hameseb6c8f52010-09-18 09:07:10 +0000580 LiveInterval *li = &lis->getInterval(*itr);
Lang Hames27601ef2008-11-16 12:12:54 +0000581
Evan Cheng90f95f82009-06-14 20:22:55 +0000582 unsigned physReg = vrm->getRegAllocPref(li->reg);
Lang Hames6699fb22009-08-06 23:32:48 +0000583
Lang Hames27601ef2008-11-16 12:12:54 +0000584 if (physReg == 0) {
585 const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
Jakob Stoklund Olesen714c0eb2011-06-16 20:37:45 +0000586 physReg = liRC->getRawAllocationOrder(*mf).front();
Lang Hames27601ef2008-11-16 12:12:54 +0000587 }
Misha Brukman2a835f92009-01-08 15:50:22 +0000588
589 vrm->assignVirt2Phys(li->reg, physReg);
Lang Hames27601ef2008-11-16 12:12:54 +0000590 }
Lang Hames27601ef2008-11-16 12:12:54 +0000591}
592
Lang Hameseb6c8f52010-09-18 09:07:10 +0000593bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
Lang Hames27601ef2008-11-16 12:12:54 +0000594
Evan Chengb1290a62008-10-02 18:29:27 +0000595 mf = &MF;
596 tm = &mf->getTarget();
597 tri = tm->getRegisterInfo();
Lang Hames27601ef2008-11-16 12:12:54 +0000598 tii = tm->getInstrInfo();
Andrew Trick16f72dd2012-02-10 04:10:26 +0000599 mri = &mf->getRegInfo();
Evan Chengb1290a62008-10-02 18:29:27 +0000600
Lang Hames27601ef2008-11-16 12:12:54 +0000601 lis = &getAnalysis<LiveIntervals>();
602 lss = &getAnalysis<LiveStacks>();
Evan Chengb1290a62008-10-02 18:29:27 +0000603 loopInfo = &getAnalysis<MachineLoopInfo>();
Lang Hames33198392010-09-02 08:27:00 +0000604 rmf = &getAnalysis<RenderMachineFunction>();
Evan Chengb1290a62008-10-02 18:29:27 +0000605
Owen Anderson49c8aa02009-03-13 05:55:11 +0000606 vrm = &getAnalysis<VirtRegMap>();
Jakob Stoklund Olesencfa81012011-11-12 23:17:52 +0000607 spiller.reset(createInlineSpiller(*this, MF, *vrm));
Evan Chengb1290a62008-10-02 18:29:27 +0000608
Jakob Stoklund Olesend9e5c762012-01-05 00:26:49 +0000609 mri->freezeReservedRegs(MF);
Lang Hames54cc2ef2010-07-19 15:22:28 +0000610
Lang Hames030c4bf2010-01-26 04:49:58 +0000611 DEBUG(dbgs() << "PBQP Register Allocating for " << mf->getFunction()->getName() << "\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000612
Evan Chengb1290a62008-10-02 18:29:27 +0000613 // Allocator main loop:
Misha Brukman2a835f92009-01-08 15:50:22 +0000614 //
Evan Chengb1290a62008-10-02 18:29:27 +0000615 // * Map current regalloc problem to a PBQP problem
616 // * Solve the PBQP problem
617 // * Map the solution back to a register allocation
618 // * Spill if necessary
Misha Brukman2a835f92009-01-08 15:50:22 +0000619 //
Evan Chengb1290a62008-10-02 18:29:27 +0000620 // This process is continued till no more spills are generated.
621
Lang Hames27601ef2008-11-16 12:12:54 +0000622 // Find the vreg intervals in need of allocation.
623 findVRegIntervalsToAlloc();
Misha Brukman2a835f92009-01-08 15:50:22 +0000624
Lang Hames20df03c2012-03-26 23:07:23 +0000625 const Function* func = mf->getFunction();
626 std::string fqn =
627 func->getParent()->getModuleIdentifier() + "." +
628 func->getName().str();
629 (void)fqn;
630
Lang Hames27601ef2008-11-16 12:12:54 +0000631 // If there are non-empty intervals allocate them using pbqp.
Lang Hameseb6c8f52010-09-18 09:07:10 +0000632 if (!vregsToAlloc.empty()) {
Evan Chengb1290a62008-10-02 18:29:27 +0000633
Lang Hames27601ef2008-11-16 12:12:54 +0000634 bool pbqpAllocComplete = false;
635 unsigned round = 0;
636
Lang Hamesab62b7e2010-10-04 12:13:07 +0000637 while (!pbqpAllocComplete) {
638 DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000639
Lang Hamesab62b7e2010-10-04 12:13:07 +0000640 std::auto_ptr<PBQPRAProblem> problem =
641 builder->build(mf, lis, loopInfo, vregsToAlloc);
Lang Hames20df03c2012-03-26 23:07:23 +0000642
643#ifndef NDEBUG
644 if (pbqpDumpGraphs) {
645 std::ostringstream rs;
646 rs << round;
647 std::string graphFileName(fqn + "." + rs.str() + ".pbqpgraph");
648 std::string tmp;
649 raw_fd_ostream os(graphFileName.c_str(), tmp);
650 DEBUG(dbgs() << "Dumping graph for round " << round << " to \""
651 << graphFileName << "\"\n");
652 problem->getGraph().dump(os);
653 }
654#endif
655
Lang Hamesab62b7e2010-10-04 12:13:07 +0000656 PBQP::Solution solution =
657 PBQP::HeuristicSolver<PBQP::Heuristics::Briggs>::solve(
658 problem->getGraph());
Lang Hames233fd9c2009-08-18 23:34:50 +0000659
Lang Hamesab62b7e2010-10-04 12:13:07 +0000660 pbqpAllocComplete = mapPBQPToRegAlloc(*problem, solution);
Lang Hames27601ef2008-11-16 12:12:54 +0000661
Lang Hamesab62b7e2010-10-04 12:13:07 +0000662 ++round;
Lang Hames27601ef2008-11-16 12:12:54 +0000663 }
Evan Chengb1290a62008-10-02 18:29:27 +0000664 }
665
Lang Hames27601ef2008-11-16 12:12:54 +0000666 // Finalise allocation, allocate empty ranges.
667 finalizeAlloc();
Evan Chengb1290a62008-10-02 18:29:27 +0000668
Lang Hamesc4bcc772010-07-20 07:41:44 +0000669 rmf->renderMachineFunction("After PBQP register allocation.", vrm);
670
Lang Hameseb6c8f52010-09-18 09:07:10 +0000671 vregsToAlloc.clear();
672 emptyIntervalVRegs.clear();
Lang Hames27601ef2008-11-16 12:12:54 +0000673
David Greene30931542010-01-05 01:25:43 +0000674 DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
Lang Hames27601ef2008-11-16 12:12:54 +0000675
Misha Brukman2a835f92009-01-08 15:50:22 +0000676 return true;
Evan Chengb1290a62008-10-02 18:29:27 +0000677}
678
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000679FunctionPass* llvm::createPBQPRegisterAllocator(
Lang Hames8d857662011-06-17 07:09:01 +0000680 std::auto_ptr<PBQPBuilder> builder,
681 char *customPassID) {
682 return new RegAllocPBQP(builder, customPassID);
Evan Chengb1290a62008-10-02 18:29:27 +0000683}
684
Lang Hamesf70e7cc2010-09-23 04:28:54 +0000685FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
686 if (pbqpCoalescing) {
687 return createPBQPRegisterAllocator(
688 std::auto_ptr<PBQPBuilder>(new PBQPBuilderWithCoalescing()));
689 } // else
690 return createPBQPRegisterAllocator(
691 std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));
Lang Hameseb6c8f52010-09-18 09:07:10 +0000692}
Evan Chengb1290a62008-10-02 18:29:27 +0000693
694#undef DEBUG_TYPE